SBC-SHAP: Increasing the Accessibility and Interpretability of Machine Learning Algorithms for Sepsis Prediction.

Abstract

Background: Sepsis is a life-threatening condition that is one of the major causes of death worldwide. Early detection of sepsis is required for fast initialization of an appropriate therapy. Complete blood count data containing information about white blood cells, platelets, hemoglobin, red blood cells, and mean corpuscular volume could serve as early indicators. However, previous approaches are limited by their interpretability (i.e., investigating the influence of feature values on individual predictions) and accessibility (i.e., easy accessibility for clinicians without programming expertise).

Methods: We developed a graph-based approach for training machine learning (ML) algorithms to incorporate time-series information for prediction based on complete blood count data. Additionally, we investigated the effect of integrating different ratios to a healthy reference measurement to improve the performance of the previously published ML model. Finally, we developed a web application based on our approaches to increase accessibility.

Results: While it was irrelevant how exactly the ratio was formed, our approach increased the sensitivity at 80% specificity across all ML models from up to 78.2% to up to 82.9% on an internal dataset (i.e., same tertiary care center) and from 65.4% to 73.4% on an external dataset (i.e., independent tertiary care center) for prospective time-series analysis. Additionally, we propose SBC-SHAP (https://mdoa-tools.bi.denbi.de/sbc-shap), a web application that visualizes the sepsis risks and individual interpretations of several ML classifiers.

Conclusions: We are confident that this tool will increase the interpretability and accessibility of ML models for predicting sepsis based on complete blood count data. SBC-SHAP is open-sourced on https://github.com/danielwalke/sbc_app.